semantic-memory adds local memory tools for Hermes agents
semantic-memory is a for AI agents and retrieval-based answer systems that need lasting memory. It keeps data on the user’s machine instead of sending it to the cloud, and stores facts, documents, and document chunks in SQLite. Search combines BM25, , and , so it can find both exact word matches and meaning-based matches.
Stored items can also have typed graph edges between them, which lets the system keep relationships between pieces of information. An MCP server is included with 18 tools, and it works with Hermes, , and Cursor. The default embedding option is Candle, which on the CPU in Rust and does not need a separate service, API key, or cloud account.
People who already run Ollama can use Ollama for , and a Mock option exists for tests.
Key points
- It includes an MCP server that can connect to Hermes.
- Memory stays local in SQLite instead of going to a cloud service.
- Search combines BM25 and for both word matches and meaning matches.
- Candle is the default embedding option and needs no API key or separate service.
- Ollama can be used for if it is already running.